Sistema de auxílio ao diagnóstico do glaucoma em imagens do fundo de olho: método baseado em transfer learning e estimação usando regressão

Abstract

Glaucoma is a chronic eye disease that affects the optic nerve, primarily caused by the gradual degeneration of the optic nerve. This disease is the leading cause of irreversible blindness worldwide. Projections from the World Health Organization suggest that this disease could affect 111.8 million people globally by 2040, particularly in populations with limited access to regular ophthalmic examinations. Due to their ability to efficiently analyze large volumes of data, computer vision techniques have contributed to the early diagnosis of diseases, including glaucoma. However, these techniques often require significant computational power. Therefore, research into methods that offer lower computational costs while maintaining adequate accuracy rates is essential for glaucoma screening in remote areas or regions with limited healthcare and communication infrastructure. Thus, the primary objective of this thesis is to propose a new method for assisting glaucoma diagnosis based on a regression model and a reduced convolutional neural network. In this context, the ACRIMA, AIROGS, LAG, and RIM-ONE databases are used and preprocessed (standardized) to center all images on the optic disc and resized to 224 x 224 x 3. A reduced VGG16 architecture is employed to optimize the extraction and dimensionality reduction of relevant features from fundus images during training. In the testing phase, a ridge regression model estimates these features from the vectorized image, preserving diagnostic accuracy while reducing computational costs. A web-based system is implemented to enable glaucoma screening through fundus image analysis, with support for portable devices, allowing usage in locations with limited infrastructure. To evaluate the performance of the glaucoma classification model, accuracy, precision, sensitivity, F1-score, and specificity metrics are used. The results indicate that convolutional models stood out in terms of performance, particularly the VGG16 (128, 512) configuration with the Minimum Redundancy Maximum Relevance (MRMR) feature selection criterion. This configuration achieved a sensitivity of 97.14% with the ACRIMA dataset and 86.54% with the RIM-ONE dataset, outperforming the original VGG16 model (512, 4096). Finally, it can be concluded that reducing the fully connected layer of neural networks also decreases computational costs without significant loss in evaluated metrics, making them suitable for classification tasks. This work's key contributions include the adaptation of selection methods to make the feature maps of a convolutional network more representative, the application of a regression model to predict the features learned by the convolutional network, and the development of a web-based system functional across different devices to assist in the automatic diagnosis of glaucoma, particularly on computationally limited devices.